Research method and coverage
What the evidence supports, and where the teaching model begins
The tutorials connect released models to small calculations a reader can inspect. Papers, release posts, model cards, and reference code establish the model-specific claims; the figures make selected relationships visible.
How claims are checked
- Read the relevant primary paper sections and official release material. Search excerpts help locate evidence; they do not substitute for the source.
- Separate the public weight release from an announcement, paper submission, or repository staging date. State uncertainty when visibility history cannot be recovered.
- Distinguish inherited methods from each release’s contribution. Stored parameters, active computation, cache memory, and measured latency are different quantities.
- Label invented dimensions and scripted data on the figure. A storage calculation does not establish model accuracy or hardware speed.
- Keep source citations beside claims and identify the relevant sections in each chapter’s source list.
A history of mechanisms
The collection runs from Llama 2 in July 2023 to releases in September 2026. It selects releases that documented a distinct technique, rather than listing every fine-tune, quantization or checkpoint. Several sizes of one release share a page when they use the same technique.
Downloadable weights do not imply published training data, training code, or unrestricted reuse. Each chapter names its weight license and the limits of reproducibility. This is a technical history, not a current model ranking or purchase recommendation.
The tests verify browser behavior, figures, calculations, and navigation. They do not rerun model training or reproduce the papers’ benchmark scores. Inaccessible sources and uncertain dates are recorded rather than filled in by inference.
Family source notes
These links open the formatted research memos on GitHub. The same Markdown files are included beside this page for local reference.